The AI Shift: Is AI supercharging science?

金融时报（科技），金融时报科技版块 - 来自金融时报的最新科技新闻与分析。关注

This article is an on-site version of our The AI Shift newsletter. Premium subscribers can sign up to get the newsletter delivered every Thursday. Standard subscribers can upgrade to Premium here, or explore all FT newsletters

Welcome back to The AI Shift, our weekly exploration of how AI is changing work and jobs. This week we’re digging into how AI tools are changing the way scientists carry out their research, off the back of a really interesting new study led by a team of researchers at Google.

John writes

Before we get into the findings, I want to make a meta-point. A lot of the economic research coming out of the big AI labs has focused very heavily on what their proprietary data tells them, which means we’ve had a glut of papers answering variations of the questions “what tasks are people using AI for?” and “what jobs are therefore most exposed to AI?” This is certainly useful, but it’s really only answering the first part of a much larger question of how AI is changing jobs and work. Whether using AI for a task makes the job as a whole less or more complex, less or more intense, whether it boosts productivity or simply reveals rigid bottlenecks — all of these require going a step further and looking at how workers’ use of AI for particular tasks fits into their job as a whole.

So what I really appreciated about this piece of research is that it did exactly that, combining data on how scientists are using Google’s LLM Gemini with a detailed survey of scientific researchers to paint a much fuller picture of how, where and why AI is and isn’t transforming the way science is carried out today.

Starting with the basics, the study found AI is already very heavily embedded in how scientific research is carried out. Almost half of surveyed US and UK scientists are using AI on a daily basis, and 80 per cent use it at least weekly. And by far the most common usage for LLMs and specialised models alike is to analyse or model research data.

But for me the most valuable part of the study was the additional survey data showing how this all maps onto the extensive range of tasks that make up the typical working week for a researcher, and what this means for the extent to which AI productivity gains and other benefits in one part of the job can spread to others and to the quantity and quality of scientific output more broadly. Data analysis is evidently where scientists are getting the most value from AI, but only about a fifth of their typical week is spent on analysing and interpreting data. So even full AI automation of those tasks would still leave the overall job essentially intact, and even dramatic productivity gains in the data analysis step might have a much more muted impact overall.

This is essentially what the rest of the survey results found. While the average surveyed scientist reported saving almost seven hours per week thanks to AI, only a portion of that time is being invested in starting additional research projects, and a significant amount goes on manually auditing and verifying the results of work performed by AI — what the researchers termed the “verification tax”. In addition, surveyed researchers reported that the biggest bottleneck for their overall progress is physical experiments and data collection, and that bottlenecks had moved downstream in the past two years — for example from carrying out exploratory research and analysis to doing physical lab experiments. As a result, 41 per cent of surveyed scientists said their backlog of untested theories had increased, suggesting AI is making it easier and faster to do the computational work that generates promising avenues for real-world research, but the physical world is not keeping pace, attenuating any speed-up in the overall scientific discovery process.

I should be clear that the overall sentiment from the surveyed researchers was that AI is proving very valuable to their work, and that they are bullish on its impact on tangible scientific output. But I found the nuances on lingering inefficiencies, bottlenecks and allocation of scientists’ time across tasks invaluable for thinking about how even hugely impressive AI capabilities and extensive usage can have much more modest impacts on final output than some might imagine.

Sarah writes

Thanks John, very interesting, and some elements — like the AI “verification tax” — resonate with what we hear from other professions. I thought I’d supplement your dive into the data with some conversations with young scientists about how this actually looks and feels at the coalface (or rather, at the lab bench).

One theme that came up was the mixed blessing of AI making it easier to pursue avenues which previously might have taken months, even years, of scientific “grunt work”. Aleksy Kwiatkowski, who has finished a PhD at the University of Oxford at the intersection of machine learning and chemistry, and is about to start a job at an AI science start-up focused on the molecules generated by fungi, told me it had become important to keep an eye on the difference between “being more productive versus feeling more productive”. There is now so much information at your fingertips, he said, that it’s important to be able to “orchestrate that and understand what’s actually most useful to you.”

Wojtek Treyde, who is wrapping up a PhD (also at Oxford) on computational drug discovery, said scientists often call this ability “taste” — an admittedly “very opaque concept” which roughly means “picking the right scientific problem to work on.” (Those of you who read our newsletter a couple of weeks ago on mathematicians will recall a similar theme.)

For early-career scientists, the increased requirement for “taste” presents something of a challenge, since they don’t yet have years of experience to draw on, but Treyde reckons younger scientists also have an advantage of their own: bravery. “I think you also need [to be] bold and courageous at the same time, right?”

Meanwhile, the scientists I spoke to said the need for real-world experimentation, flagged in the Google research, helps to explain why AI hasn’t yet lived up to Silicon Valley’s breathless claims that it will “solve” all diseases. One bottleneck in the physical world? Mice. “You can’t make mice do things faster,” said Leah Morris, executive director of an organisation called Encode: AI for Science. “You can’t make them grow faster. You can’t have the virus replicate faster . . . And then the other piece is regulatory — we still have drug trials in terms of how to get things through and make them safe for humans. There are people working on how to do this more efficiently . . . but as it stands, there’s a lot that AI cannot touch.”

AI can still be transformational, though. Murray Cox, who works on cancer immunotherapies and is in the final year of a PhD at Imperial College London, told me it was important to distinguish between general-purpose LLMs and specialised models (which might not be LLMs at all) that can be built to develop a “latent understanding of data which we as scientists don’t have, or have never previously had.” Used in this way, AI “opens up a whole new door to different types of discoveries and different types of experiments”. But this often requires AI engineers to go and work in scientific labs, where “they’re able to write whole new models to develop a new layer of understanding,” as Cox put it.

And this raises another human problem: bringing the right people together. Encode: AI for Science, which Cox is also involved with, takes AI experts and plugs them into UK-based science labs (supported by UK government funding). They have also created an open-access collection of research problems proposed by academics where AI expertise is needed, in an attempt to address this “matching” bottleneck. “We had so many scientists asking us for AI talent, [and] AI talent saying, ‘Hey, I want to work on something cooler than chatbots. I want to work on something more interesting than the next tool to help your kids cheat on their homework’,” Morris explained.

Kwiatkowski and Treyde, meanwhile, are part of a group of early-career computational scientists called CompMotifs, who convene workshops and hackathons on how to build better tools for science.

In sum, the bottlenecks are real, but people are busy working on them too. And at a time when we’re hearing a lot about AI’s destructive potential, it’s heartening to remember there are a lot of smart people out there who are working thoughtfully and collaboratively to make the best of this technology. As Treyde put it: “It would be foolish to pretend there are no risks associated with AI . . . but I think at the same time this technology has huge potential right? We also should not pretend that this progress isn’t happening . . . I think it’s sort of on us to make sure that this is being used for the right things.”

The AI Shift: Is AI supercharging science?

John writes

Sarah writes

Recommended reading

Recommended newsletters for you